Frontier AI Is Becoming a State-Licensed Weapon

Since July 1, Fable 5 is once again available worldwide. The press led with relief: Anthropic got its most powerful model back, Washington backed down, case closed. In fact, the model's return is not the story. The story is the mechanism that return just validated.

In three weeks, a government cut off a frontier AI, filtered access to it, then switched it back on under its own terms. And it kept its hand on the switch.

A two-line recap

For anyone who missed it: on June 12, the U.S. Department of Commerce, the agency that runs export controls in the United States, barred access to Fable 5 and Mythos 5 for any foreign national, citing national security. Anthropic, unable to sort users by nationality on a global API, cut everything off. We covered that flip elsewhere, triggered by Amazon, its lead backer. On June 30, the controls were lifted.

That is where the story worth telling begins.

The price of the return

The reversal came at a price, and Anthropic paid it. In his June 30 letter, Commerce Secretary Howard Lutnick lays out the terms: no more export license required, in exchange for firm commitments. The company must proactively detect and address the security risks of its models, work with the government on standards for future models, Mythos and Fable as well as what comes next, and report any malicious activity.

One clause ties the whole thing together: Lutnick explicitly reserves the right to reassess and adjust the licensing requirements if circumstances change. The switch that shut Fable off on a Friday night stayed in place, now wrapped in a contract.

A detail that says a lot about the balance of power: the negotiation was not led by CEO Dario Amodei, who spent much of the year on frosty terms with the administration, but by co-founder Tom Brown. You send in the one who still gets a hearing.

A model that came back diminished

The model that returns has changed. Fable 5 reappears throttled, with new guardrails, and some cyber tasks, even plain coding, take a short-term hit. As for Mythos 5, the least restricted version, it stays reserved for around a hundred U.S. organizations vetted by the government: federal agencies and hand-picked private companies. For users in the rest of the world, Mythos stays closed.

For the developers who had built workflows on the unrestricted model, the downgrade is not theoretical: a capability that worked in May can simply refuse in July, with no version to fall back to.

So the state did more than flip the model back on. It decided who gets what, at what capability, on which territory. A new power, exercised live on a global market.

The rationale collapsed, the power stayed

The most revealing part is the reason for the reversal. Commerce reopened access on an unexpected ground: Anthropic's own tests showed that rival models reproduced the exact exploit that had triggered the ban. The flaw existed elsewhere, available more or less everywhere, which hollowed out the national security argument.

The justification dissolved. The control apparatus stayed intact, and even settled into a signed contract. That is how a precedent works: the pretext fades, the power infrastructure remains.

A framework, not an accident

An isolated case would read like a regulatory mishap. Two other pieces say otherwise.

On June 2, a presidential executive order set up a framework in which developers voluntarily submit their frontier models to the government, up to thirty days before release, ahead of any distribution to "trusted partners." The text takes care to rule out any mandatory license. On paper, no one is forcing anything. In practice, Commerce had just placed the country's most advanced model under a de facto license.

And Anthropic is not alone in this corridor. On June 26, OpenAI shipped its GPT-5.6 generation in limited access to about twenty partners, also government-vetted, at the government's request. The company noted that it did not believe "this kind of government access process should become the norm." The phrasing mostly says it is becoming exactly that.

The dependency cuts both ways

This shift reaches well beyond U.S. borders. European governments were weighing whether to anchor critical systems to these models. They are discovering that a foreign vendor can be unplugged, filtered or conditioned by an administration they have no leverage over. That is precisely the question raised by Europe's sovereign large-model project, and the one running through the debate around the AI Act: who actually controls the AI you rely on.

Frontier AI thus joins a category once thought reserved for weapons or breakthrough military technology: strategic assets whose tap a state controls. You no longer really buy them; you access them under conditions, revisable ones.

Fable 5 is back, and that is exactly what makes the episode instructive. The whole world is building its tools on a handful of models that a single government can now, legally and in a matter of minutes, switch off.